Image Super-Resolution via Deep Feature Recalibration Network

Recent years have witnessed remarkable progress in convolutional neural network (CNN) based image super-solution (SR) methods. Existing methods tend to deepen the network by means of residual skip connections to achieve better performance. However, these methods are still hard to be applied in real-world applications due to the requirement of its heavy computation. In this paper, we propose a Deep Feature Recalibration Network (DFRN), which strives for efficiency yet effective networks. We divide the process of network nonlinear mapping into two steps: information integration and feature enhancement, and proposed two types of block models: Multi-Scale Information Integration Block (MSIIB) and Feature Recalibration Block (FRB). MSIIB integrates the representation of the input data in the network with different size of receptive fields. FRB enhances the information via obtaining the attention along two different dimensions (channel and plane space of feature maps) respectively. By combining MSIIB and FRB, we provide a more efficient and time-saving method for SISR. Experiments show that the proposed DFRN method outperforms state-of-the-art methods in terms of both objective evaluation metrics (PSNR, SSIM, and running speed) and subjective perception on the generated images.